dsputil_mmx.c

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

      "pxor %%mm7,%%mm7\n"	/* mm7 holds the sum */
      "1:\n"
      "movq (%0),%%mm1\n"	/* mm1 = pix1[0-7] */
      "movq (%1),%%mm2\n"	/* mm2 = pix2[0-7] */
      "movq 8(%0),%%mm3\n"	/* mm3 = pix1[8-15] */
      "movq 8(%1),%%mm4\n"	/* mm4 = pix2[8-15] */

      /* todo: mm1-mm2, mm3-mm4 */
      /* algo: substract mm1 from mm2 with saturation and vice versa */
      /*       OR the results to get absolute difference */
      "movq %%mm1,%%mm5\n"
      "movq %%mm3,%%mm6\n"
      "psubusb %%mm2,%%mm1\n"
      "psubusb %%mm4,%%mm3\n"
      "psubusb %%mm5,%%mm2\n"
      "psubusb %%mm6,%%mm4\n"

      "por %%mm1,%%mm2\n"
      "por %%mm3,%%mm4\n"

      /* now convert to 16-bit vectors so we can square them */
      "movq %%mm2,%%mm1\n"
      "movq %%mm4,%%mm3\n"

      "punpckhbw %%mm0,%%mm2\n"
      "punpckhbw %%mm0,%%mm4\n"
      "punpcklbw %%mm0,%%mm1\n"	/* mm1 now spread over (mm1,mm2) */
      "punpcklbw %%mm0,%%mm3\n"	/* mm4 now spread over (mm3,mm4) */

      "pmaddwd %%mm2,%%mm2\n"
      "pmaddwd %%mm4,%%mm4\n"
      "pmaddwd %%mm1,%%mm1\n"
      "pmaddwd %%mm3,%%mm3\n"

      "add %3,%0\n"
      "add %3,%1\n"

      "paddd %%mm2,%%mm1\n"
      "paddd %%mm4,%%mm3\n"
      "paddd %%mm1,%%mm7\n"
      "paddd %%mm3,%%mm7\n"

      "decl %%ecx\n"
      "jnz 1b\n"

      "movq %%mm7,%%mm1\n"
      "psrlq $32, %%mm7\n"	/* shift hi dword to lo */
      "paddd %%mm7,%%mm1\n"
      "movd %%mm1,%2\n"
      : "+r" (pix1), "+r" (pix2), "=r"(tmp) 
      : "r" ((long)line_size) , "m" (h)
      : "%ecx");
    return tmp;
}

static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) {
    int tmp;
  asm volatile (
      "movl %3,%%ecx\n"
      "pxor %%mm7,%%mm7\n"
      "pxor %%mm6,%%mm6\n"
      
      "movq (%0),%%mm0\n"
      "movq %%mm0, %%mm1\n"
      "psllq $8, %%mm0\n"
      "psrlq $8, %%mm1\n"
      "psrlq $8, %%mm0\n"
      "movq %%mm0, %%mm2\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm0\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm2\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm0\n"
      "psubw %%mm3, %%mm2\n"
      
      "add %2,%0\n"
      
      "movq (%0),%%mm4\n"
      "movq %%mm4, %%mm1\n"
      "psllq $8, %%mm4\n"
      "psrlq $8, %%mm1\n"
      "psrlq $8, %%mm4\n"
      "movq %%mm4, %%mm5\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm4\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm5\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm4\n"
      "psubw %%mm3, %%mm5\n"
      "psubw %%mm4, %%mm0\n"
      "psubw %%mm5, %%mm2\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm0, %%mm3\n\t"
      "pcmpgtw %%mm2, %%mm1\n\t"
      "pxor %%mm3, %%mm0\n"
      "pxor %%mm1, %%mm2\n"
      "psubw %%mm3, %%mm0\n" 
      "psubw %%mm1, %%mm2\n"
      "paddw %%mm0, %%mm2\n"
      "paddw %%mm2, %%mm6\n"

      "add %2,%0\n"
      "1:\n"
  
      "movq (%0),%%mm0\n"
      "movq %%mm0, %%mm1\n"
      "psllq $8, %%mm0\n"
      "psrlq $8, %%mm1\n"
      "psrlq $8, %%mm0\n"
      "movq %%mm0, %%mm2\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm0\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm2\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm0\n"
      "psubw %%mm3, %%mm2\n"
      "psubw %%mm0, %%mm4\n"
      "psubw %%mm2, %%mm5\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm4, %%mm3\n\t"
      "pcmpgtw %%mm5, %%mm1\n\t"
      "pxor %%mm3, %%mm4\n"
      "pxor %%mm1, %%mm5\n"
      "psubw %%mm3, %%mm4\n" 
      "psubw %%mm1, %%mm5\n"
      "paddw %%mm4, %%mm5\n"
      "paddw %%mm5, %%mm6\n"
      
      "add %2,%0\n"
      
      "movq (%0),%%mm4\n"
      "movq %%mm4, %%mm1\n"
      "psllq $8, %%mm4\n"
      "psrlq $8, %%mm1\n"
      "psrlq $8, %%mm4\n"
      "movq %%mm4, %%mm5\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm4\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm5\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm4\n"
      "psubw %%mm3, %%mm5\n"
      "psubw %%mm4, %%mm0\n"
      "psubw %%mm5, %%mm2\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm0, %%mm3\n\t"
      "pcmpgtw %%mm2, %%mm1\n\t"
      "pxor %%mm3, %%mm0\n"
      "pxor %%mm1, %%mm2\n"
      "psubw %%mm3, %%mm0\n" 
      "psubw %%mm1, %%mm2\n"
      "paddw %%mm0, %%mm2\n"
      "paddw %%mm2, %%mm6\n"

      "add %2,%0\n"
      "subl $2, %%ecx\n"
      " jnz 1b\n"

      "movq %%mm6, %%mm0\n"
      "punpcklwd %%mm7,%%mm0\n"
      "punpckhwd %%mm7,%%mm6\n"
      "paddd %%mm0, %%mm6\n"
      
      "movq %%mm6,%%mm0\n"
      "psrlq $32, %%mm6\n"
      "paddd %%mm6,%%mm0\n"
      "movd %%mm0,%1\n"
      : "+r" (pix1), "=r"(tmp) 
      : "r" ((long)line_size) , "g" (h-2)
      : "%ecx");
      return tmp;
}

static int hf_noise16_mmx(uint8_t * pix1, int line_size, int h) {
    int tmp;
    uint8_t * pix= pix1;
  asm volatile (
      "movl %3,%%ecx\n"
      "pxor %%mm7,%%mm7\n"
      "pxor %%mm6,%%mm6\n"
      
      "movq (%0),%%mm0\n"
      "movq 1(%0),%%mm1\n"
      "movq %%mm0, %%mm2\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm0\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm2\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm0\n"
      "psubw %%mm3, %%mm2\n"
      
      "add %2,%0\n"
      
      "movq (%0),%%mm4\n"
      "movq 1(%0),%%mm1\n"
      "movq %%mm4, %%mm5\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm4\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm5\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm4\n"
      "psubw %%mm3, %%mm5\n"
      "psubw %%mm4, %%mm0\n"
      "psubw %%mm5, %%mm2\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm0, %%mm3\n\t"
      "pcmpgtw %%mm2, %%mm1\n\t"
      "pxor %%mm3, %%mm0\n"
      "pxor %%mm1, %%mm2\n"
      "psubw %%mm3, %%mm0\n" 
      "psubw %%mm1, %%mm2\n"
      "paddw %%mm0, %%mm2\n"
      "paddw %%mm2, %%mm6\n"

      "add %2,%0\n"
      "1:\n"
  
      "movq (%0),%%mm0\n"
      "movq 1(%0),%%mm1\n"
      "movq %%mm0, %%mm2\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm0\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm2\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm0\n"
      "psubw %%mm3, %%mm2\n"
      "psubw %%mm0, %%mm4\n"
      "psubw %%mm2, %%mm5\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm4, %%mm3\n\t"
      "pcmpgtw %%mm5, %%mm1\n\t"
      "pxor %%mm3, %%mm4\n"
      "pxor %%mm1, %%mm5\n"
      "psubw %%mm3, %%mm4\n"
      "psubw %%mm1, %%mm5\n"
      "paddw %%mm4, %%mm5\n"
      "paddw %%mm5, %%mm6\n"
      
      "add %2,%0\n"
      
      "movq (%0),%%mm4\n"
      "movq 1(%0),%%mm1\n"
      "movq %%mm4, %%mm5\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm4\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm5\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm4\n"
      "psubw %%mm3, %%mm5\n"
      "psubw %%mm4, %%mm0\n"
      "psubw %%mm5, %%mm2\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm0, %%mm3\n\t"
      "pcmpgtw %%mm2, %%mm1\n\t"
      "pxor %%mm3, %%mm0\n"
      "pxor %%mm1, %%mm2\n"
      "psubw %%mm3, %%mm0\n" 
      "psubw %%mm1, %%mm2\n"
      "paddw %%mm0, %%mm2\n"
      "paddw %%mm2, %%mm6\n"

      "add %2,%0\n"
      "subl $2, %%ecx\n"
      " jnz 1b\n"

      "movq %%mm6, %%mm0\n"
      "punpcklwd %%mm7,%%mm0\n"
      "punpckhwd %%mm7,%%mm6\n"
      "paddd %%mm0, %%mm6\n"
      
      "movq %%mm6,%%mm0\n"
      "psrlq $32, %%mm6\n"
      "paddd %%mm6,%%mm0\n"
      "movd %%mm0,%1\n"
      : "+r" (pix1), "=r"(tmp) 
      : "r" ((long)line_size) , "g" (h-2)
      : "%ecx");
      return tmp + hf_noise8_mmx(pix+8, line_size, h);
}

static int nsse16_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
    MpegEncContext *c = p;
    int score1= sse16_mmx(c, pix1, pix2, line_size, h);
    int score2= hf_noise16_mmx(pix1, line_size, h) - hf_noise16_mmx(pix2, line_size, h);

    if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
    else  return score1 + ABS(score2)*8;
}

static int nsse8_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
    MpegEncContext *c = p;
    int score1= sse8_mmx(c, pix1, pix2, line_size, h);
    int score2= hf_noise8_mmx(pix1, line_size, h) - hf_noise8_mmx(pix2, line_size, h);

    if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
    else  return score1 + ABS(score2)*8;
}

static int vsad_intra16_mmx(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) {
    int tmp;
    
    assert( (((int)pix) & 7) == 0);
    assert((line_size &7) ==0);
    
#define SUM(in0, in1, out0, out1) \
      "movq (%0), %%mm2\n"\
      "movq 8(%0), %%mm3\n"\
      "add %2,%0\n"\
      "movq %%mm2, " #out0 "\n"\
      "movq %%mm3, " #out1 "\n"\
      "psubusb " #in0 ", %%mm2\n"\
      "psubusb " #in1 ", %%mm3\n"\
      "psubusb " #out0 ", " #in0 "\n"\
      "psubusb " #out1 ", " #in1 "\n"\
      "por %%mm2, " #in0 "\n"\
      "por %%mm3, " #in1 "\n"\
      "movq " #in0 ", %%mm2\n"\
      "movq " #in1 ", %%mm3\n"\
      "punpcklbw %%mm7, " #in0 "\n"\
      "punpcklbw %%mm7, " #in1 "\n"\
      "punpckhbw %%mm7, %%mm2\n"\
      "punpckhbw %%mm7, %%mm3\n"\
      "paddw " #in1 ", " #in0 "\n"\
      "paddw %%mm3, %%mm2\n"\
      "paddw %%mm2, " #in0 "\n"\
      "paddw " #in0 ", %%mm6\n"

    
  asm volatile (
      "movl %3,%%ecx\n"
      "pxor %%mm6,%%mm6\n"
      "pxor %%mm7,%%mm7\n"
      "movq (%0),%%mm0\n"
      "movq 8(%0),%%mm1\n"
      "add %2,%0\n"
      "subl $2, %%ecx\n"
      SUM(%%mm0, %%mm1, %%mm4, %%mm5)
      "1:\n"
      
      SUM(%%mm4, %%mm5, %%mm0, %%mm1)
      
      SUM(%%mm0, %%mm1, %%mm4, %%mm5)
      
      "subl $2, %%ecx\n"
      "jnz 1b\n"

      "movq %%mm6,%%mm0\n"
      "psrlq $32, %%mm6\n"
      "paddw %%mm6,%%mm0\n"
      "movq %%mm0,%%mm6\n"
      "psrlq $16, %%mm0\n"
      "paddw %%mm6,%%mm0\n"
      "movd %%mm0,%1\n"
      : "+r" (pix), "=r"(tmp) 
      : "r" ((long)line_size) , "m" (h)
      : "%ecx");
    return tmp & 0xFFFF;
}
#undef SUM

static int vsad_intra16_mmx2(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) {
    int tmp;
    
    assert( (((int)pix) & 7) == 0);
    assert((line_size &7) ==0);
    
#define SUM(in0, in1, out0, out1) \
      "movq (%0), " #out0 "\n"\
      "movq 8(%0), " #out1 "\n"\
      "add %2,%0\n"\
      "psadbw " #out0 ", " #in0 "\n"\
      "psadbw " #out1 ", " #in1 "\n"\
      "paddw " #in1 ", " #in0 "\n"\
      "paddw " #in0 ", %%mm6\n"

  asm volatile (
      "movl %3,%%ecx\n"
      "pxor %%mm6,%%mm6\n"
      "pxor %%mm7,%%mm7\n"
      "movq (%0),%%mm0\n"
      "movq 8(%0),%%mm1\n"